Summary

• PhysicsX and GB1 are combining Large Geometry Models and Large Physics Models to generate yacht hull designs and predict their hydrodynamic performance.
• PhysicsX reports five times fewer high-fidelity RANS simulations, 100 times greater accuracy against GB1's previous panel method and 100 times faster evaluations.
• The results are company-reported engineering comparisons, not independently verified racing gains. The development extends an existing partnership into the design requirements for the 2029 America's Cup.

British engineering company PhysicsX is developing a generative AI system that creates racing yacht hull geometries and predicts their performance, allowing Britain’s GB1 sailing team to investigate designs beyond those available through conventional parameter-based optimisation.

The companies presented the new workflow at Nvidia GTC Berlin on 21 October, extending their existing engineering partnership towards the rules governing the 2029 America’s Cup.

PhysicsX reports that the approach reduces requirements for high-fidelity fluid dynamics simulations by a factor of five, delivers evaluations 100 times faster and improves predictive accuracy by a factor of 100 compared with GB1’s previous panel method.

Those figures refer to different engineering comparisons and remain company-reported. They do not establish that a completed yacht has achieved a corresponding improvement in speed or that the method will deliver the same advantages across unrelated engineering applications.

Generative models widen the design space

Racing yacht design involves evaluating the interaction between hull geometry, water movement and the forces acting on the vessel. Subtle changes to a hull can influence resistance and stability, while the overall performance of the boat also depends on its other components and the conditions encountered during competition.

Engineers traditionally use computer-aided design and numerical simulation to investigate alternative shapes. Optimisation software can adjust a set of dimensions or design parameters, compare the results and repeat the process to identify promising configurations.

However, the possibilities examined are partly determined by the parameters defined at the beginning. Designs outside that representation may not be considered, even where a different geometric arrangement could offer better performance.

PhysicsX and GB1 are developing Large Geometry Models that generate new three-dimensional shapes from representations learned during training.

The models are pretrained on geometric information and then adapted to the particular range of shapes and requirements relevant to GB1’s programme.

Instead of changing only a limited collection of predefined dimensions, the system can propose new hull geometries within the constraints established by the engineering team.

Large Physics Models then estimate the performance of those candidates, allowing engineers to compare many alternatives before selecting particular designs for more detailed investigation.

The approach can increase the number of possibilities examined within a fixed development period, although a larger search does not guarantee that the best proposed geometry will satisfy every manufacturing or competition requirement.

Engineers must still assess whether a promising shape can be manufactured, integrated with the rest of the boat and operated reliably under the conditions anticipated during racing.

Physics models reduce the simulation workload

Computational fluid dynamics provides a way to examine how water moves around a hull, including the forces generated by changing flow conditions.

High-fidelity simulations based on the Reynolds-averaged Navier–Stokes equations, commonly described as RANS, can provide detailed engineering information but require substantial computing resources.

Running such calculations for every candidate geometry would restrict the number of designs that could be investigated within the time available.

PhysicsX is using AI predictions to evaluate candidates quickly and concentrate more demanding simulation work on designs that appear promising or particularly informative.

The company reports a fivefold reduction in high-fidelity RANS simulation requirements during its design process.

That does not mean an individual RANS calculation has become five times faster. The improvement concerns how often the workflow needs to rely on those more demanding simulations.

PhysicsX separately claims that its AI predictions are 100 times more accurate than GB1’s previous panel-method simulation and that evaluations can be completed 100 times faster overall.

The accuracy comparison is relative to the earlier panel method used by the team, rather than every available conventional fluid dynamics technique.

The release does not provide a complete independent benchmark identifying the absolute errors, representative test cases and measurement definitions behind the comparisons.

Without that information, the reported results support the company’s account of its development progress but cannot establish universal performance advantages.

Detailed numerical simulation and physical testing remain important because AI models can become less reliable when proposed geometries or operating conditions differ substantially from their training data.

Active learning directs further exploration

The workflow connects geometry generation and performance prediction through an active learning system.

As candidate designs are evaluated, the resulting information helps determine which parts of the design space the models should explore next.

Rather than examining possible shapes uniformly, the software can direct computing resources towards regions where further investigation may reveal useful performance improvements.

The technique is particularly applicable when a design problem contains many interacting variables and only a limited number of high-fidelity calculations can be completed.

Nevertheless, the effectiveness of active learning depends on the criteria used to select candidates and on the reliability of predictions made by the underlying models.

If the system repeatedly favours designs because its performance model contains an error, further exploration could reinforce a misleading assumption rather than identify a genuine improvement.

Engineers therefore need methods for testing predictions against more reliable calculations and physical measurements.

PhysicsX uses Nvidia GPUs for distributed model training, supported by CUDA-X libraries including cuBLAS and cuFFT.

The company is also evaluating Nvidia Warp for its wider physics AI software environment.

These tools provide computing resources and numerical capabilities, but they do not remove the need for suitable training information or independent engineering validation.

New rules create a different engineering problem

GB1’s existing partnership with PhysicsX was established for the 2027 America’s Cup programme, with work focused on accelerating engineering analysis and optimisation within a defined design space.

The latest collaboration is intended to address the rules being introduced for the 2029 competition, which the partners say create opportunities to explore different hull geometries.

Changes in competition requirements can alter which engineering solutions are useful. A shape optimised under one set of limits may be unsuitable or less competitive when the design framework changes.

Generating candidates more freely could therefore help the team investigate options that would be difficult to discover through a narrow parameter-based process.

However, the rules, manufacturing requirements and remaining engineering work will determine which concepts can become finished racing equipment.

The companies have not disclosed the resulting hull design or published independently verified performance comparisons from a completed boat.

Any competitive advantage must ultimately be established through the performance of the integrated vessel rather than the accuracy or speed of an individual design calculation.

Engineering applications extend beyond sailing

The underlying methods have potential applications in industries that depend on repeated physical simulations.

Automotive engineers examine aerodynamic performance, cooling and structural behaviour, while aerospace manufacturers model airflow and mechanical loads across complex components.

Energy and industrial equipment developers must evaluate similar relationships between geometry, material properties and operating conditions.

Generative models could help those teams create design alternatives, with physics models identifying candidates that warrant closer investigation.

Commercial adoption would nevertheless require relevant training data, integration with established engineering software and evidence that predictions remain reliable for each intended application.

A model developed for the hydrodynamic behaviour of racing yachts cannot simply be transferred to semiconductor devices or aircraft structures without substantial adaptation.

Organisations would also need to compare the cost of developing and maintaining AI models against the computing resources and engineering time saved.

Where a conventional simulation is already inexpensive, training a specialised model may offer limited economic benefit. More demanding design programmes may create a stronger case, particularly when the model can be reused across many evaluations.

PhysicsX’s work with GB1 demonstrates an attempt to change the search process itself, rather than only accelerating existing calculations.

The next stage will involve validating whether the models consistently identify useful geometries and whether their predicted advantages survive detailed simulation, manufacture and physical testing.